On the Design and Analysis of Data Center Network Architectures for Interconnecting Dual-Port Servers. Dawei Li and Prof. Jie Wu Temple University

Transcription

1 On the Design and Analysis of Data Center Network Architectures for Interconnecting Dual-Port Servers Dawei Li and Prof. Jie Wu Temple University

2 Outline Introduction Preliminaries Maximizing the number of dual-port servers given network diameter and switch port number SWCube SWKautz Related Existing architectures On the Comparison of Various Architectures Evaluation of SWCube and SWKautz Conclusion

3 Introduction The number of servers in modern and future data centers will tend to be very large. Challenge: how to design network architectures to interconnect large numbers of servers, and also to meet the requirements of Data Center Networks (DCN). Basic connections in a DCN: Server-server Server-switch (switch-server) Switch-switch

4 Introduction DCN architecture classification: based on whether the interconnection intelligence is put on switches or servers. Switch-centric Server-centric Server-centric More than two Network Interface Cards (NICs) are used: BCube, DCell No more than two NICs are used: FiConn, HCN&BCN, Dpillar.

5 Introduction Main contributions: We propose the concept of Normalized Switch Delay (NSD), denoted by c, to unify the design and analysis of DCNs for dual-port servers. We ask the following fundamental question: what is the maximum number of dual-port servers that any architecture can accommodate at most, given network diameter d, and switch port number n? And give an upper bound on this maximum number. We propose two novel DCN architectures that try to achieve this upper bound. We also show that the two proposed architectures have good properties for DCNs.

6 Preliminaries Some definitions: A hop is a path, from one node to another node of the same kind, which consists of no other nodes of the same kind. Thus, we have switch-to-switch hops and server-to-server hops. Server-to-server hops consist of server-to-server-direct hops and server-to-server-via-a-switch hops. The length of a path between two servers is the number of server-to-server-direct hop(s), plus 1 + c times the number of server-to-server-via-switch hop(s) in the path. The distance of two servers is the length of the shortest path between the two servers. The diameter of a DCN architecture is the maximum distance among all pairs of servers.

7 Preliminaries A preview of the influence of NSD (c).

8 Maximizing the Number of Dual-Port Servers Given Network Diameter and Switch Port Number

9 Maximizing the Number of Dual-Port Servers Given Network Diameter and Switch Port Number Consider a DCN architecture with dual-port servers when c = 0.

10 Maximizing the Number of Dual-Port Servers Given Network Diameter and Switch Port Number Proof of Theorem 1

11 Maximizing the Number of Dual-Port Servers Given Network Diameter and Switch Port Number when c!= 0.

12 Maximizing the Number of Dual-Port Servers Given Network Diameter and Switch Port Number The upper bound may not be achievable

13 Maximizing the Number of Dual-Port Servers Given Network Diameter and Switch Port Number

14 SWCube The Generalized Hypercube A node W is represented by a k-tuple Two nodes are connected directly by a link if and only if their addresses differ at one bit. SWCube Construction 1.) replace the nodes in the original generalized hypercube with switches 2.) insert one server into each link that connects two switches

15 SWCube

16 SWCube Properties Lemma 1: The distance of two servers that are along the same dimension is at most 2. Lemma 2: The distance of two servers that are not along the same dimension is at most k + 1. Theorem 3: The diameter of an SWCube(r, k) is d = k+1. Theorem 4: In terms of network diameter and switch port number, the number of servers in an SWCube(r; k) is

17 SWKautz

18 SWKautz SWKautz construction 1.) replace each node in the original KA(n/2, k) graph with an n-port switch. 2.) remove the direction of all the edges and insert a server into each edge.

19 SWKautz

20 Existing architectures FiConn

21 Existing architectures HCN & BCN

22 Existing architectures DPillar

23 On the Comparison of Various Architectures

24 On the Comparison of Various Architectures The Number of Servers Given d and n

25 On the Comparison of Various Architectures Hardware Interconnection Cost per Server

26 On the Comparison of Various Architectures Influence of c on Various Architectures

27 Evaluation of SWCube and SWKautz Routing Properties of SWCube and SWKautz SWCube

28 Evaluation of SWCube and SWKautz Routing Properties of SWCube and SWKautz SWKautz Conclusion: both SWCube and SWKautz have good fault-tolerance properties.

29 Evaluation of SWCube and SWKautz

30 Evaluation of SWCube and SWKautz Routing Simulation With Congestion SWCube(13,2) 2028 servers SWKautz(12,2) 1872 servers Conclusion: Both SWCube and SWKautz have the capability of efficiently handling network congestion.

31 Conclusion We propose the concept of Normalized Switch Delay (NSD), denoted by c, to unify the design and analysis of DCNs for dual-port servers. We ask the following fundamental question: what is the maximum number of dual-port servers that any architecture can accommodate at most, given network diameter d, and switch port number n? And give an upper bound on this maximum number. We propose two novel DCN architectures that try to achieve this upper bound. Comparisons with the existing one demonstrate various advantages. Evaluations on themselves show they have good properties for DCNs.

32